Giskard Scan Results

The data are filtered by ethnicity to detect ethical biases in the landmark detection model.

Issues

1 major

(Cached (300W) with ethnicity) filtered using 'ethnicity: asian' NME_mean = 6.285 (Global = 3.949) +59.16% than global 29 samples affected
(15.4% of dataset) Show details Hide details

Examples

The data are filtered by head pose to detect biases in the landmark detection model.

Issues

3 major

(Cached (300W) with head-pose) filtered using 'head pose: 1e-05 < roll < inf' NME_mean = 5.050 (Global = 3.903) +29.37% than global 116 samples affected
(61.7% of dataset) Show details Hide details

Examples

(Cached (300W) with head-pose) filtered using 'head pose: -90 < yaw < -45' NME_mean = 6.097 (Global = 3.757) +62.27% than global 18 samples affected
(9.6% of dataset) Show details Hide details

Examples

(Cached (300W) with head-pose) filtered using 'head pose: -inf < pitch < -1e-05' NME_mean = 4.340 (Global = 3.860) +12.42% than global 161 samples affected
(85.6% of dataset) Show details Hide details

Examples

Images from the dataset are blurred, recolored and resized to test the robustness of the model to transformations.

Issues

1 medium

300W resizing with ratios: 0.5 NME_mean = 3.936 (Global = 3.845) +2.37% than global 188 samples affected
(100.0% of dataset) Show details Hide details